Files
Jyotisha/frontend/src/lib/agent-observability.ts
T
Jesse_Chen ff70ba87b0 feat(consult): deliver the route's strict method with the evidence instead of listing 1592 filenames
Activating the skill returned 129,651 bytes, of which 99KB was a flat list of
1,592 undifferentiated file paths against 30KB of actual method. The one line
telling the model to open the strict-workflow router sat inside that method,
so no reference was ever opened and every answer was composed from the model's
own background knowledge over server evidence.

The route is already decided server-side and the skill already states which
checklist each route requires, so the selection needs no model turn: read the
mandated sections from the hash-pinned package and hand them to the model with
the evidence they apply to. A route the router declares no checklist for is
reported as such rather than filled in with another route's.

The receipt now reports delivered sections separately from model-initiated
reads, because only one of those is under the model's control.

Co-authored-by: Cursor <cursoragent@cursor.com>
2026-08-18 20:53:07 +08:00

197 lines
6.7 KiB
TypeScript

import { z } from "zod";
import { consultationDomainSchema } from "./consultation-domain-registry.ts";
/**
* Closed, non-PII observability contract for Agent runs.
*
* There is intentionally no free-form metadata bag. Raw input/output text,
* prompts, messages, birth data, names, email addresses, secrets, API keys,
* provider payloads, stack traces and internal filesystem paths are not fields
* in this schema. Every object is strict, so unknown fields fail closed.
*/
const opaqueIdSchema = z.string()
.min(1)
.max(128)
.regex(/^[A-Za-z0-9][A-Za-z0-9._:-]*$/, "invalid opaque identifier");
const machineValueSchema = z.string()
.min(1)
.max(128)
.regex(/^[A-Za-z0-9][A-Za-z0-9._:+-]*$/, "invalid machine value");
const machineCodeSchema = z.string()
.min(1)
.max(80)
.regex(/^[a-z][a-z0-9._-]*$/, "invalid machine code");
const durationMsSchema = z.number().int().min(0).max(7 * 24 * 60 * 60 * 1000);
const countSchema = z.number().int().min(0).max(1_000_000);
const tokenCountSchema = z.number().int().min(0).max(1_000_000_000);
export const agentObservabilityStepStatuses = [
"completed",
"failed",
"cancelled",
"skipped",
] as const;
export const billingSettlementResults = [
"completed",
"cancelled",
"failed",
"not_applicable",
] as const;
export type AgentBillingSettlementResult = (typeof billingSettlementResults)[number];
export type AgentSettlementResult = Exclude<AgentBillingSettlementResult, "not_applicable">;
/**
* Closed vocabulary for why the model stopped stepping. These are the provider
* finish reasons plus the two Mastra adds; anything else normalizes to
* `unknown` so an unrecognized provider string can never become a log field.
*
* `tool-calls` on a run that produced no answer means the step budget ran out
* while the model still wanted to call a tool, which is otherwise only
* inferable from the recorded step list.
*/
export const agentModelFinishReasons = [
"stop",
"length",
"content-filter",
"tool-calls",
"error",
"other",
"tripwire",
"retry",
"unknown",
] as const;
export type AgentModelFinishReason = (typeof agentModelFinishReasons)[number];
const knownModelFinishReasons = new Set<string>(agentModelFinishReasons);
export function toAgentModelFinishReason(value: unknown): AgentModelFinishReason {
return typeof value === "string" && knownModelFinishReasons.has(value)
? value as AgentModelFinishReason
: "unknown";
}
export type AgentSettlementTelemetryOutcome = Readonly<{
billingSettlementResult: AgentSettlementResult;
errorCode?: string;
}>;
export function settlementTelemetryOutcome(
settlementResult: AgentSettlementResult,
errorCode?: string,
): AgentSettlementTelemetryOutcome {
if (settlementResult === "failed") {
return {
billingSettlementResult: "failed",
errorCode: "settlement_failed",
};
}
return {
billingSettlementResult: settlementResult,
...(errorCode === undefined ? {} : { errorCode }),
};
}
export const agentObservabilityToolCallSchema = z.object({
name: machineCodeSchema,
durationMs: durationMsSchema,
status: z.enum(agentObservabilityStepStatuses),
// Closed classification of why the call failed. Never upstream error text.
failureCode: machineCodeSchema.optional(),
}).strict().readonly();
export const agentObservabilityContractPhaseSchema = z.object({
phase: machineCodeSchema,
durationMs: durationMsSchema.optional(),
status: z.enum(agentObservabilityStepStatuses),
}).strict().readonly();
export const agentObservabilityEventSchema = z.object({
runId: opaqueIdSchema.optional(),
requestId: opaqueIdSchema.optional(),
sessionId: opaqueIdSchema.optional(),
caseId: opaqueIdSchema.optional(),
agentVersion: machineValueSchema.optional(),
skillVersion: machineValueSchema.optional(),
modelVersion: machineValueSchema.optional(),
policyVersion: machineValueSchema.optional(),
toolCalls: z.array(agentObservabilityToolCallSchema).max(64).optional(),
contractPhases: z.array(agentObservabilityContractPhaseSchema).max(64).optional(),
retryCount: z.number().int().min(0).max(100).optional(),
errorCode: machineCodeSchema.optional(),
// Why the model stopped, and how many model steps the run consumed across
// every attempt. Both are enum-like machine values, never provider text.
modelFinishReason: z.enum(agentModelFinishReasons).optional(),
modelStepCount: countSchema.optional(),
// How many reference documents the model opened after loading the skill, and how many strict-method
// sections the server delivered with the evidence. Both are needed to read the other: zero reads is
// only a gap in the answer's method if nothing was delivered either.
skillReferenceReads: countSchema.optional(),
methodologySections: countSchema.optional(),
inputTokens: tokenCountSchema.optional(),
outputTokens: tokenCountSchema.optional(),
evidenceCount: countSchema.optional(),
claimCount: countSchema.optional(),
sectionCount: countSchema.optional(),
themeCoverage: z.array(consultationDomainSchema).max(10).optional(),
reportJobDurationMs: durationMsSchema.optional(),
reportJobPeakMemoryBytes: z.number().int().min(0).max(Number.MAX_SAFE_INTEGER).optional(),
billingSettlementResult: z.enum(billingSettlementResults).optional(),
}).strict().refine(
(event) => Boolean(event.runId || event.requestId || event.sessionId || event.caseId),
{ message: "at least one controlled identifier is required" },
).readonly();
export type AgentObservabilityEvent = z.infer<typeof agentObservabilityEventSchema>;
export type AgentObservabilitySink = (event: AgentObservabilityEvent) => void;
export type AgentObservabilityLogger = (event: unknown) => AgentObservabilityEvent;
function consoleSink(event: AgentObservabilityEvent): void {
console.info("[agent-observability]", JSON.stringify(event));
}
export function createAgentObservabilityLogger(
sink: AgentObservabilitySink = consoleSink,
): AgentObservabilityLogger {
return (event) => {
const parsed = agentObservabilityEventSchema.parse(event);
try {
sink(parsed);
} catch {
// Observability transport failure must not change the business response.
}
return parsed;
};
}
export const logAgentObservability = createAgentObservabilityLogger();
const knownErrorCodes = new Set([
"runtime_contract_incomplete",
"empty_answer",
"calculation_failed",
"timeout",
"cancelled",
"settlement_failed",
]);
export function toAgentObservabilityErrorCode(error: unknown): string {
if (error instanceof DOMException) {
if (error.name === "TimeoutError") return "timeout";
if (error.name === "AbortError") return "cancelled";
}
if (error instanceof Error && knownErrorCodes.has(error.message)) return error.message;
return "calculation_failed";
}